A neural-network approach for predicting time to cardiovascular diseases in HIV patients based on real-world data

At the end of 2021, 38.4 million people were living with HIV (PLWH) worldwide. The advent of anti retroviral therapy (ART) has significantly reduced the mortality and increased life expectancy of PLWH. Nowadays, the management of people with HIV on virological suppression is partly focused on the onset of comorbidities, such as the occurrence of cardiovascular diseases (CVDs). In this real-world study, we analyse the 15 years CVD risk in PLWH, following a survival analysis approach based on neural networks (NNs). We adopt a NN-based deep learning approach to flexibly model and predict the time to a CVD event, relaxing the linearity and the proportional-hazard assumptions typical of the COX model and including time-varying features. Results of this approach are compared to the ones obtained via more classical survival analysis methods, both in terms of predictive performance and interpretability. A further aim is to explore the potential of deep learning approaches in modelling survival data with time-varying features for supporting decision-making in real clinical setting.